d/numam-dpdk
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

bond: add mode 5

Browse Source 
Add support for mode 5 (Transmit load balancing) into pmd driver

This patch add support for Adaptive transmit load balancing (mode 5) to the
librte_pmd_bond library. This mode provides an adaptive transmit load
balancing. It dynamically changes the transmitting slave, according to the
computed load.

Further details are described here:
https://www.kernel.org/doc/Documentation/networking/bonding.txt
In implementation callback is used for sorting slave order - providing
statistics for burst function about slave bandwith usage  and sort
interfaces due to usage.

Difference in this implementation vs Linux implementation:
- We Are trying send all pkts – If one interface hasn’t send packets we are
trying to send rest of packets by other slaves sorted previously by callback
function.

Some implementation details:
- Every 100ms is taken obytes statistics from every slave.
- Every 10 ms the slaves in  table are sorted and updated by callback -
bandwidth and successfully transmitted bytes from previous iteration which
happens every 100 ms
- There is callback function which updates this statistics for transparency and
for rather intensive computation involved in this mode.

Test report: http://dpdk.org/ml/archives/dev/2014-November/008729.html

Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
Tested-by: SunX Jiajia <sunx.jiajia@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
This commit is contained in:
Daniel Mrzyglod 2014-11-27 16:33:40 +00:00 committed by Thomas Monjalon
parent 46fb436836
commit 7c76a747e6
4 changed files with 167 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
lib/librte_pmd_bond/rte_eth_bond.h
View File

@ -96,6 +96,12 @@ extern "C" {
* to rx_burst should be at least 2 times the slave count size.
*
*/
#define BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING (5)
/**< Adaptive TLB (Mode 5)
* This mode provides an adaptive transmit load balancing. It dynamically
* changes the transmitting slave, according to the computed load. Statistics
* are collected in 100ms intervals and scheduled every 10ms */
/* Balance Mode Transmit Policies */
#define BALANCE_XMIT_POLICY_LAYER2 (0)
/**< Layer 2 (Ethernet MAC) */

1
lib/librte_pmd_bond/rte_eth_bond_args.c
View File

@ -174,6 +174,7 @@ bond_ethdev_parse_slave_mode_kvarg(const char *key __rte_unused,
case BONDING_MODE_BROADCAST:
#endif
case BONDING_MODE_8023AD:
case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
return 0;
default:
RTE_BOND_LOG(ERR, "Invalid slave mode value (%s) specified", value);

161
lib/librte_pmd_bond/rte_eth_bond_pmd.c
View File

@ -30,7 +30,7 @@
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <stdlib.h>
#include <rte_mbuf.h>
#include <rte_malloc.h>
#include <rte_ethdev.h>
@ -41,11 +41,16 @@
#include <rte_kvargs.h>
#include <rte_dev.h>
#include <rte_alarm.h>
#include <rte_cycles.h>
#include "rte_eth_bond.h"
#include "rte_eth_bond_private.h"
#include "rte_eth_bond_8023ad_private.h"
#define REORDER_PERIOD_MS 10
/* Table for statistics in mode 5 TLB */
static uint64_t tlb_last_obytets[RTE_MAX_ETHPORTS];
static uint16_t
bond_ethdev_rx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
{
@ -361,6 +366,144 @@ xmit_slave_hash(const struct rte_mbuf *buf, uint8_t slave_count, uint8_t policy)
return hash % slave_count;
}
struct bwg_slave {
uint64_t bwg_left_int;
uint64_t bwg_left_remainder;
uint8_t slave;
};
static int
bandwidth_cmp(const void *a, const void *b)
{
const struct bwg_slave *bwg_a = a;
const struct bwg_slave *bwg_b = b;
int64_t diff = (int64_t)bwg_b->bwg_left_int - (int64_t)bwg_a->bwg_left_int;
int64_t diff2 = (int64_t)bwg_b->bwg_left_remainder -
(int64_t)bwg_a->bwg_left_remainder;
if (diff > 0)
return 1;
else if (diff < 0)
return -1;
else if (diff2 > 0)
return 1;
else if (diff2 < 0)
return -1;
else
return 0;
}
static void
bandwidth_left(int port_id, uint64_t load, uint8_t update_idx,
struct bwg_slave *bwg_slave)
{
struct rte_eth_link link_status;
rte_eth_link_get(port_id, &link_status);
uint64_t link_bwg = link_status.link_speed * 1000000ULL / 8;
if (link_bwg == 0)
return;
link_bwg = (link_bwg * (update_idx+1) * REORDER_PERIOD_MS);
bwg_slave->bwg_left_int = (link_bwg - 1000*load) / link_bwg;
bwg_slave->bwg_left_remainder = (link_bwg - 1000*load) % link_bwg;
}
static void
bond_ethdev_update_tlb_slave_cb(void *arg)
{
struct bond_dev_private *internals = arg;
struct rte_eth_stats slave_stats;
struct bwg_slave bwg_array[RTE_MAX_ETHPORTS];
uint8_t slave_count;
uint64_t tx_bytes;
uint8_t update_stats = 0;
uint8_t i, slave_id;
internals->slave_update_idx++;
if (internals->slave_update_idx >= REORDER_PERIOD_MS)
update_stats = 1;
for (i = 0; i < internals->active_slave_count; i++) {
slave_id = internals->active_slaves[i];
rte_eth_stats_get(slave_id, &slave_stats);
tx_bytes = slave_stats.obytes - tlb_last_obytets[slave_id];
bandwidth_left(slave_id, tx_bytes,
internals->slave_update_idx, &bwg_array[i]);
bwg_array[i].slave = slave_id;
if (update_stats)
tlb_last_obytets[slave_id] = slave_stats.obytes;
}
if (update_stats == 1)
internals->slave_update_idx = 0;
slave_count = i;
qsort(bwg_array, slave_count, sizeof(bwg_array[0]), bandwidth_cmp);
for (i = 0; i < slave_count; i++)
internals->active_slaves[i] = bwg_array[i].slave;
rte_eal_alarm_set(REORDER_PERIOD_MS * 1000, bond_ethdev_update_tlb_slave_cb,
(struct bond_dev_private *)internals);
}
static uint16_t
bond_ethdev_tx_burst_tlb(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
{
struct bond_tx_queue *bd_tx_q = (struct bond_tx_queue *)queue;
struct bond_dev_private *internals = bd_tx_q->dev_private;
struct rte_eth_dev *primary_port =
&rte_eth_devices[internals->primary_port];
uint16_t num_tx_total = 0;
uint8_t i, j;
uint8_t num_of_slaves = internals->active_slave_count;
uint8_t slaves[RTE_MAX_ETHPORTS];
struct ether_hdr *ether_hdr;
struct ether_addr primary_slave_addr;
struct ether_addr active_slave_addr;
if (num_of_slaves < 1)
return num_tx_total;
memcpy(slaves, internals->active_slaves,
sizeof(internals->active_slaves[0]) * num_of_slaves);
ether_addr_copy(primary_port->data->mac_addrs, &primary_slave_addr);
if (nb_pkts > 3) {
for (i = 0; i < 3; i++)
rte_prefetch0(rte_pktmbuf_mtod(bufs[i], void*));
}
for (i = 0; i < num_of_slaves; i++) {
ether_addr_copy(&internals->slaves[slaves[i]].persisted_mac_addr,
&active_slave_addr);
for (j = num_tx_total; j < nb_pkts; j++) {
if (j + 3 < nb_pkts)
rte_prefetch0(rte_pktmbuf_mtod(bufs[j+3], void*));
ether_hdr = rte_pktmbuf_mtod(bufs[j], struct ether_hdr *);
if (is_same_ether_addr(&ether_hdr->s_addr, &primary_slave_addr))
ether_addr_copy(&active_slave_addr, &ether_hdr->s_addr);
}
num_tx_total += rte_eth_tx_burst(slaves[i], bd_tx_q->queue_id,
bufs + num_tx_total, nb_pkts - num_tx_total);
if (num_tx_total == nb_pkts)
break;
}
return num_tx_total;
}
static uint16_t
bond_ethdev_tx_burst_balance(void *queue, struct rte_mbuf **bufs,
uint16_t nb_pkts)
@ -690,6 +833,7 @@ mac_address_slaves_update(struct rte_eth_dev *bonded_eth_dev)
bond_mode_8023ad_mac_address_update(bonded_eth_dev);
break;
case BONDING_MODE_ACTIVE_BACKUP:
case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
default:
for (i = 0; i < internals->slave_count; i++) {
if (internals->slaves[i].port_id ==
@ -751,6 +895,10 @@ bond_ethdev_mode_set(struct rte_eth_dev *eth_dev, int mode)
"Using mode 4, it is necessary to do TX burst and RX burst "
"at least every 100ms.");
break;
case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
eth_dev->tx_pkt_burst = bond_ethdev_tx_burst_tlb;
eth_dev->rx_pkt_burst = bond_ethdev_rx_burst_active_backup;
break;
default:
return -1;
}
@ -876,7 +1024,7 @@ slave_add(struct bond_dev_private *internals,
}
slave_details->link_status_wait_to_complete = 0;
/* clean tlb_last_obytes when adding port for bonding device */
memcpy(&(slave_details->persisted_mac_addr), slave_eth_dev->data->mac_addrs,
sizeof(struct ether_addr));
}
@ -965,6 +1113,9 @@ bond_ethdev_start(struct rte_eth_dev *eth_dev)
if (internals->mode == BONDING_MODE_8023AD)
bond_mode_8023ad_start(eth_dev);
if (internals->mode == BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING)
bond_ethdev_update_tlb_slave_cb(internals);
return 0;
}
@ -994,6 +1145,10 @@ bond_ethdev_stop(struct rte_eth_dev *eth_dev)
}
}
if (internals->mode == BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING) {
rte_eal_alarm_cancel(bond_ethdev_update_tlb_slave_cb, internals);
}
internals->active_slave_count = 0;
internals->link_status_polling_enabled = 0;
@ -1242,6 +1397,7 @@ bond_ethdev_promiscuous_enable(struct rte_eth_dev *eth_dev)
break;
/* Promiscuous mode is propagated only to primary slave */
case BONDING_MODE_ACTIVE_BACKUP:
case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
default:
rte_eth_promiscuous_enable(internals->current_primary_port);
}
@ -1270,6 +1426,7 @@ bond_ethdev_promiscuous_disable(struct rte_eth_dev *dev)
break;
/* Promiscuous mode is propagated only to primary slave */
case BONDING_MODE_ACTIVE_BACKUP:
case BONDING_MODE_ADAPTIVE_TRANSMIT_LOAD_BALANCING:
default:
rte_eth_promiscuous_disable(internals->current_primary_port);
}

2
lib/librte_pmd_bond/rte_eth_bond_private.h
View File

@ -104,7 +104,6 @@ struct bond_slave_details {
uint8_t link_status_poll_enabled;
uint8_t link_status_wait_to_complete;
uint8_t last_link_status;
/**< Port Id of slave eth_dev */
struct ether_addr persisted_mac_addr;
};
@ -149,6 +148,7 @@ struct bond_dev_private {
struct mode8023ad_private mode4;
struct rte_kvargs *kvlist;
uint8_t slave_update_idx;
};
extern struct eth_dev_ops default_dev_ops;